You are on page 1of 11

Computers and Geotechnics 122 (2020) 103499

Contents lists available at ScienceDirect

Computers and Geotechnics


journal homepage: www.elsevier.com/locate/compgeo

Research Paper

Inverse analysis of Hypoplastic Clay model for computing deformations T


caused by excavations
Sangrae Kim, Richard J. Finno

Department of Civil and Environmental Engineering, Northwestern University, Evanston, IL 60208, USA

ARTICLE INFO ABSTRACT

Keywords: For effective applications of the adaptive management approach, inverse analysis procedures are used to update
Inverse analysis model parameters based on observations made during early stages of construction projects. These updated
Optimization parameters are then used to compute the responses at later stages. This paper describes the inverse analysis of
Small-strain stiffness lateral wall deformations from a deep supported excavation in Chicago using the Hypoplastic Clay model. Initial
Deep supported excavation
parameters for the soil model are based on inverse analyses of laboratory test results and site-specific correla-
Lateral wall deformation
tions with in situ data of Chicago clays. Field observations made during construction are used to conduct sen-
sitivity and correlation analyses, and the most effective parameters to optimize are selected. These parameters
are optimized at three depths during excavation where plane strain conditions were applicable. Changes in the
optimized parameters at the excavation levels are analyzed to assess how and if the parameters are optimized
satisfactorily. Computed responses using the optimized parameters are compared to field measurements, and
recommendations regarding the parameters to optimize are provided.

1. Introduction responses at more critical later stages of construction.


For effective inverse analysis procedures, the best parameters to
Minimizing damage to adjacent structures and utilities is an essen- optimize are selected based on site-specific observations made at an
tial goal when making deep excavations in urban environments. For this excavation site. The excavation must have reliable performance data
purpose, one must carefully evaluate the ground responses during the with respect to construction progress. Simpson Querry Biomedical
entire construction process. To make design predictions and to properly Research Center (SQBRC) is one such case study where reliable field
assess the performance during construction and make reasonable pre- measurements were made for a deep supported excavation in down-
dictions of movements at later construction stages, the use of finite town Chicago. Detailed analyses of the observed performance for the
element modeling (FEM) has steadily increased and has been applied to excavation were presented by Finno et al. [19]. Performance data ob-
various excavation projects in urban areas [1–11]. When properly tained from the SQBRC site are used for inverse analysis herein.
executed, these finite element simulations can produce computed Kim and Finno [18] conducted inverse analyses using the measured
ground responses which reflect the overall performance during ex- lateral wall deformations of the excavation site and showed that small-
cavations. strain stiffness responses need to be considered to allow meaningful
Efforts to properly monitor, simulate and predict performance updates of parameters during early stages of excavation. To illustrate an
during construction have led to the enhancement of the classical ob- effective procedure for selecting parameters to optimize and obtaining
servational approach [12], allowing a sequence of measurement and reasonably optimized soil parameters of small-strain stiffness models,
prediction updates in near-real-time. This enhancement, called “adap- the Hypoplastic Clay (HC) model [20] is used to represent the behavior
tive management,” can be used to predict, monitor and control ground of the site clays for the inverse analyses presented herein.
deformations that develop during geotechnical constructions Initial model parameters must be defined based on available data,
[5,13–18]. After data sets (e.g., lateral deformations, settlements, strut i.e., site-specific laboratory/field data or other data sets in the same
loads) are collected during early stages of a project, soil parameters are geological unit. This paper describes procedures to select initial para-
optimized through inverse analysis procedures. These updated, or op- meter values of the HC model using previous inverse analysis studies of
timized, parameters form the basis of new simulations to predict laboratory test results and site-specific correlations based on in situ


Corresponding author.
E-mail addresses: Rae.Kim@terracon.com (S. Kim), r-finno@northwestern.edu (R.J. Finno).

https://doi.org/10.1016/j.compgeo.2020.103499
Received 30 July 2019; Received in revised form 16 February 2020; Accepted 16 February 2020
Available online 25 February 2020
0266-352X/ © 2020 Elsevier Ltd. All rights reserved.
S. Kim and R.J. Finno Computers and Geotechnics 122 (2020) 103499

investigations of Chicago clays. The Composite Scaled Sensitivity (CSS) is used herein to quantify
Performance data collected at the SQBRC excavation are used to the relative importance with respect to the observations of the model
conduct sensitivity and correlation analyses of the HC model para- parameters being simultaneously estimated, and is defined as [22]:
meters and, on the basis of these results, the most effective parameters 1 2
2
for optimization can be identified. Selection is evaluated in light of the ND
yi'
sensitivities and correlation coefficients of model parameters and de- CSSj = bj wii1 2
ND
bj
formation levels of the observed measurements used in the analyses.
j=1 (1)
The optimized parameters are compared with the reported range of where yi' th th
is the i computed value, bj is the j estimated parameter,
values from previous studies on various clays to show the optimization yi' bj is the sensitivity of the ith computed value with respect to the jth
process leads to reasonable values of the parameters. Lateral wall de- parameter, wii is the weight of the ith observation, and ND is the number
formations are computed at three excavated levels using both the initial of observations. Weights of the observations were calculated objectively
and optimized parameters and are compared to the observed mea- using the variance for the 95% confidence interval for the repeatability
surements. The efficacy of the HC model for the use of inverse analysis ( ± 1.5 mm/32 m) of the Shape Acceleration Array (SAA) measure-
in deep supported excavations is evaluated based on the fit between the ments. As a result, computed weights increased with depth, indicating
“predicted” and measured lateral wall movements of latter excavation the largest weight at the bottom of the SAA. A relatively large CSS value
stages, and recommendations regarding the parameters to optimize are for a parameter indicates the measurements used in the optimization
provided. are sensitive to that parameter.
However, parameters cannot be simultaneously optimized when
2. Technical background they are closely related to each other. To assess the correlations be-
tween the model parameters, correlation analyses are carried out and
2.1. Inverse Analysis results are defined in terms of the Parameter Correlation Coefficient,
PCC [22]:
Fig. 1 shows a schematic of the inverse analysis procedure used
herein. Computations of lateral wall deformations are carried out with cov (i, j)
PCC =
PLAXIS 2D 2017. A gradient approach based on the Gauss-Newton var (i )1 2 var (j )1 2
(2)
method, found in the optimization toolbox in MATLAB, is used to op-
timize the soil parameters. A Python code [17] was written to transfer where cov(i,j) is the covariance between parameters i and j, and var(i) is
data between PLAXIS and MATLAB such that the analysis procedure the variance of parameter i. Absolute values of PCC greater than 0.9
proceeds uninterrupted. Optimization was considered to be complete indicate high correlations between two parameters, i and j, and may be
when the objective function and soil parameter values changed less problematic when the two parameters are simultaneously optimized.
than 0.1% between iterations. Details of the inverse analysis procedure Finno and Calvello [5] and Rechea [21] used 0.95 as the maximum
can be found in Finno and Calvello [5] and Rechea [21]. values of the PCC to select a parameter for optimization. In this paper,
an even smaller value of 0.90 was used to help ensure that parameters
2.2. Sensitivity and correlation analyses can be simultaneously optimized and be unique. Ideally, one selects
parameters for optimization based on a combination of the highest CSS
Given the limited number of observations associated with each values and lowest PCC values.
optimization stage, optimization of all parameters of a soil model may
result in unreasonable parameter values due to the correlations and 2.3. Hypoplastic Clay model
dependencies between soil parameters. For effective and reasonable
optimizations of multiple parameters, those that have low correlations, When applying the adaptive management approach to deep sup-
but are highly sensitive to the outcome of interest must be selected in a ported excavations, it is most beneficial to optimize parameters in early
rational manner. Sensitivity and correlation analyses are used herein to stages of the project when deformation levels are relatively small.
quantitatively accomplish this task. Consequently, models that include the capability to represent small-

Fig. 1. Inverse analysis flowchart.

2
S. Kim and R.J. Finno Computers and Geotechnics 122 (2020) 103499

Table 1 Table 2
HC model parameters and their definitions. Construction stages.
Soil Parameters Definition Construction Activity Excavation Level

Large Strain λ* Slope of isotropic normal compression line Sheet pile wall installation –
κ* Slope of isotropic unloading line Excavation to elev. 0 m CCD 1
N Position of isotropic normal compression line First level tieback at elev. +0.6 m CCD –
φc Critical state friction angle Excavation to elev. −2.1 m CCD 2
νpp Stiffness fitting parameter Strut installation at elev. −1.2 m CCD –
Excavation to elev. −6.7 m CCD 3
Small Strain Ag Stress dependency of Gvh on mean normal stress
Second level tieback at elev. −6.1 m CCD –
ng Stress dependency of Gvh on mean normal stress
Excavation to elev. −8.8 m CCD 4
αg, αE, αν Anisotropy coefficient of shear, Young’s and
Poisson’s moduli
mrat Very small strain behavior upon strain path
reversals i.e., at excavation levels that are uniform across the entire site before
Rmax Size of elastic range supports at that level are installed. Lateral wall movements at Ex-
βr Material constant controlling the rate of cavation Levels 2, 3 and 4 are used as the observation for the inverse
evolution of intergranular strain tensor
χ Material constant controlling the rate of
analysis. Movements observed at Level 1 were not considered because
degradation of shear stiffness lateral wall movements were negligible, less than 3 mm in the medium
clay [18].
Structural parameters used in the finite element analyses for the
strain stiffness response are desirable [18]. A soil model that includes lateral support system at the SQBRC excavation are listed in Table 3.
such capabilities, the HC model, is selected in this work for this pur- The struts and unbonded zones of the tiebacks were modeled with
pose. point-to-point anchors, which are two-noded elastic spring elements
The HC model developed by Masin [20] uses basic principles of that transmit both tensile and compressive forces. Bonded zones of the
critical state soil mechanics to represent the behavior of clays following tiebacks were modeled with embedded beam rows, a simplified ap-
different stress paths. The soil parameters are based on effective stress proach to deal with three-dimensional stress states and deformation
and their definitions are summarized in Table 1. The reference HC patterns in a 2D plane strain model.
model [23] that defines behavior at large strains is defined by five soil Interface stiffness factors between the soil and tiebacks were com-
model parameters, λ*, κ*, N, φc′ and νpp. By incorporating the inter- puted as a function of the diameter (Deq) and horizontal spacing
granular strain concept [24] through nine parameters, Ag, ng, αG, αE, αν, (Lspacing) of the embedded beams as [25]:
mrat, Rmax, βr and χ, Masin [20] extended the reference HC model to
capture the non-linear soil stiffness at small strains. Lspacing
0.75
ISFRS = ISFRN = 2.5
Deq (3)
3. Optimization method
0.75
Lspacing
3.1. Finite element procedures ISFKF = 25
Deq (4)
The SQBRC site is located in downtown Chicago. The basement
where ISFRS, ISFRN, and ISFKF represent the axial skin stiffness factor,
excavation was 13.1 m deep with 3 levels of lateral supports as shown
lateral skin stiffness factor, and the pile base stiffness factor, respec-
in Fig. 2. Descriptions of the basic soil properties at the site are found in
tively. The embedded beam diameters were 0.15 m and the embedded
Finno et al. [19]. Ground water was encountered at elevation (elev.)
beam spacings were 1.4 m.
0 m Chicago City Datum (CCD). The excavation sequence to the nom-
The interface strength parameter, Rinter, was set as 0.67, based on
inal full excavation depth (elev. −8.8 m CCD) can be organized into 8
results of previous FEM studies of tieback-supported deep excavations
construction stages as noted in Table 2. All computations were made
[11,26].
assuming undrained conditions in the clay strata. Inverse analyses were
Optimizations were carried out for Excavation Levels 2, 3 and 4
carried out at the stages when plane strain conditions could be applied,
when plane strain conditions applied to the excavation. The model was
first optimized starting at Level 2 using the horizontal movements
versus depth collected by SAA-1 when the excavation was at Level 2. It
was recalibrated at Level 3 using the horizontal movements versus
depth collected by SAA-1 when the excavation was at Levels 2 and 3,
and at Level 4 using the SAA-1 data collected when the excavation was
at Levels 2, 3 and 4.

3.2. Field measurements

To measure the lateral deformations along the sheet pile walls, SAAs
were installed at various locations around the site. The data were col-
lected autonomously at specified time intervals and were transferred to
a project web site for immediate viewing. These SAAs were installed
directly behind the sheet pile walls and initialized after the sheet piles
were installed, but before excavation began. SAA-1 was chosen to be
used for the inverse analysis because it properly represented plane
strain conditions at the center of the north wall. Detailed analyses of the
Fig. 2. Earth retention system along north wall. field measurements during excavation can be found in Finno et al. [19].

3
S. Kim and R.J. Finno Computers and Geotechnics 122 (2020) 103499

Table 3
Model parameters for supports.
Structural Parameters Sheet Pile Wall 1st Level Tieback 2nd Level Tieback Strut

H3707 Unbonded Zone Bonded Zone Unbonded Zone Bonded Zone Pile

Type of element Plate Point-to-point anchor Embedded beam row Point-to-point anchor Embedded beam row Point-to-point anchor
Bending stiffness, EI (kN·m2/m) 4,426,000 – – – – –
Axial stiffness, EA (kN/m) 1,848,000 – – – – –
Element thickness (m) 0.71 – – – – –
Poisson’s ratio, ν 0.2 – – – – –
Young’s modulus, E (kN/m2) – – 7,995,000 – 6,675,000 –
Axial stiffness, EA (kN) – 146,000 – 219,000 – 6,985,000
Max. skin friction, Tskin,max (kN/m) – – 120.6 – 131.1 –

3.3. Initial HC model parameters Table 5


Initial HC model parameters.
Based on performance analyses of deep supported excavations in Soil Parameters Soil Strata
Chicago, it is known that the medium and stiff clay layers shown in
Fig. 2 have the biggest influence on the lateral wall deformations during Medium Clay Stiff Clay
excavations [7,19,27] and these strata are the focus of the inverse
Large Strain N 0.76 0.64
analyses presented herein. Soil parameters for the fill, beach sand and λ* 0.06 0.05
hard clay have been reasonably estimated using the HS model κ* 0.012 0.012
[5,19,21]. HS model parameters for the urban fill, beach sand and hard φ′c 33 35
clay shown in Table 4 were taken from the parameters used to suc- νpp 0.15 0.15

cessfully replicate the lateral wall movements at the Lurie excavation Small Strain Ag 4100 11,000
[21], which is adjacent to the SQBRC excavation. The equality of E50ref ng 0.6 0.6
αg, αE, αν 1.1 1.1
and Eoedref for the Fill and Beach Sand layers is consistent with experi-
mrat 1.0 1.0
mental data for granular soils [28]. Rmax 5.0 × 10−05 2.4 × 10−05
Only the medium and stiff clay layers were represented by the HC βr 0.18 0.18
model. The initial set of HC model parameters of the medium and stiff χ 1.3 1.3
clays are listed in Table 5. Sarabia [29] developed a relationship be-
tween the critical state friction angle (φc’) and the natural water content
using laboratory testing data on soft to medium stiff Chicago clay =5
(8)
samples collected from various locations and depths [30]:
Initial values of parameters N, λ* and κ* for the medium clay were
sin( c' ) 0.7385 0.9384wn (5) computed as 0.76, 0.06 and 0.012, respectively.
Initial values of φc′ for the medium and stiff clays were computed as Arboleda-Monsalve [33] used site-specific field investigations from
33° and 35°, respectively, on the basis of the natural water contents at the One Museum Park West (OMPW) project in Chicago and proposed
the SQBRC site. correlations between the N, λ* and κ* parameters and the natural water
To obtain values of N, λ* and κ*, Sarabia [29] used one-dimensional content for the stiff clay layer.
constrained compression (oedometer) tests of medium clay specimens N = 3.65wn 0.05 (9)
cut from block samples collected from two excavations made in Chi-
cago, the Block 37 [31] and Ford Engineering Center [32] excavations, = 0.34wn 0.01 (10)
and developed regression equations for these parameters with the
natural water content (wn): =4
(11)
3
N = 2.444wn + 0.1478 + log ' Initial values of N, λ* and κ* for the stiff clay were computed as
2(1 sin cv ) +1 (6) 0.64, 0.05 and 0.012, respectively.
= 0.211wn + 0.008 (7) The term νpp is a fitting parameter related to the bulk-to-shear
modulus ratio, r, from earlier versions of the HC model by the following
correlation presented by Masin [20]:
Table 4
Parameters for HS model parameters. 3r ( + ) 4
pp =
Soil Parameters Soil Strata
6r ( + )+4 (12)

Fill Beach Sand Hard Clay


Sarabia [29] determined the parameter r as 0.424, which was op-
timized using results of K0-consolidated undrained triaxial tests. Values
Type Drained Drained Undrained of λ* and κ* were obtained through regression analyses of natural water
E50ref (kPa) 13,500 48,000 200,000 contents as previously presented. Calculations using Eq. (12) showed
Eoedref (kPa) 13,500 48,000 140,000
Eurref (kPa) 40,500 144,000 600,000
similar νpp values for the medium and stiff clays, thus a value of 0.15
m 0.5 0.5 0.6 was used for νpp parameters of both clay layers.
cref (kPa) 19 0.2 0.2 Masin [20] proposed an equation for the stress dependency of the
φ (°) 30 35 35 initial shear stiffness as:
ψ (°) 2 5 0
OCR 1 1.1 2.5 ng
p'
Gvh = pr Ag
Note: pref = 100 kPa.
pr (13)

4
S. Kim and R.J. Finno Computers and Geotechnics 122 (2020) 103499

where pr is the reference pressure parameter, and Gvh is the shear


modulus at very small strains in which the double indices refer to the
propagation direction and polarization of a shear wave, respectively.
Arboleda-Monsalve et al. [34] computed Ag and ng values using la-
boratory data collected with bender elements during the Kc-re-
compression stage of triaxial tests from Block 37 [32,35] and the shear
velocities measured at the OMPW and Ford Center projects [36]. Initial
values of parameters Ag and ng of the medium clay were computed as
4100 and 0.6, respectively. The initial ng parameter for the stiff clay was
kept constant as 0.6. The Ag parameter for the stiff clay was altered to fit
the measured shear velocities with depth and was computed as 11,000.
Kim and Finno [35] showed that αG varies between 1.1 and 1.2 for
compressible Chicago clay, representing its minor degree of cross-ani-
sotropy. Values αE and αν were both computed as 1.1 using correlations
proposed by Masin and Rott [37]. A value of 1.1 was used for all three
anisotropy coefficients, αG, αE and αν.
The intergranular strain parameter, mrat, controls the very small
strain behavior upon strain path reversals. It was assumed herein that
the very small strain stiffness is controlled with parameters Ag and ng,
thus, mrat was assumed a material-independent constant of 1.0.
Parameters Rmax, βr and χ determine the size of the elastic range in
the strain space and the rate of degradation of the shear stiffness. The
intergranular strain parameter, Rmax, of the medium clay was assumed a
value of 5 × 10−5 because Masin [38] reported values of Rmax that
range from 2 × 10−5–1 × 10−4 for six different clays. Arboleda-
Monsalve [33] modified the Rmax parameter of the stiff clay to capture
in finite element simulations the observed movements at the OMPW
excavation. Arboleda-Monsalve’s modified value of Rmax, 2.4 × 10−5,
Fig. 3. Composite Scaled Sensitivities of HC model parameters.
was used as the initial parameter for the stiff clay herein.
Sarabia [29] determined the βr and χ parameters through optimi-
zations between numerical simulations and test results of undrained recompression response. The small-strain parameters have larger CSS
triaxial compression and extension tests on block samples cut from the values than the large-strain parameters for all cases. Note that sensi-
Block 37 and Ford Center excavations. Results of Sarabia’s analyses tivities of the large-strain parameters in the medium clay at Level 2 are
indicated βr and χ parameter values of 0.18 and 1.3, respectively, for close to zero.
both medium and stiff clays. Fig. 4 shows the correlations between two HC model parameters at
each optimization level for both the medium and stiff clay stratum. As
3.4. Results of statistical analyses noted previously, absolute PCC values greater than 0.9 indicate a re-
latively high correlation between two parameters and may be proble-
Soil deformations are small in excavation projects where limiting matic when the two parameters are simultaneously optimized.
ground movements is a design objective, thus failure in the affected soil Among the five deformation parameters of the HC model considered
mass is expected to be limited and deformation parameters are the most for optimization, one parameter governing small-strain and one gov-
important in these types of problems. Of all the HC model parameters erning large-strain were chosen for optimization. Sensitivity analyses
shown in Table 5, λ*, κ* and νpp were selected as candidates for para- indicated that Rmax and κ* should be the two parameters to optimize.
meters to optimize with respect to the large-strain response, i.e., However, values of κ* can be reasonably evaluated by results of con-
parameters that primarily govern soil’s deformability as opposed to ventional oedometer tests of which there are plenty of available data for
shear strength. For the small-strain response, parameters Rmax and χ Chicago clays and did not need to be additionally altered via optimi-
were selected because these parameters govern the shape of the shear zation. Furthermore, νpp had the second largest CSS of the large-strain
stiffness degradation curve and are difficult to obtain through con- parameters, and thus was selected as the large-strain parameter to be
ventional regressions with laboratory test data. Sensitivity and corre- optimized. Absolute PCC values for the two parameters, Rmax and νpp,
lation analyses were carried out to assess which of these soil parameters were lower than 0.9, indicating that the two parameters can be si-
should be optimized for the most effective inverse analysis in the multaneously optimized.
SQBRC case. As shown in Fig. 3, the lateral wall deformations were significantly
Fig. 3 shows the computed CSS values for HC model parameters of more sensitive to the small-strain parameters than the large-strain
the medium and stiff clay layers at Level 2 based on the horizontal parameters. Therefore, in addition to optimizing both the small-strain
movements versus depth collected by SAA-1 when the excavation was and large-strain parameters, only the small-strain parameter was opti-
at Level 2, at Level 3 using the horizontal movements versus depth mized to evaluate if the single parameter approach yielded better re-
when the excavation was at Levels 2 and 3, and at Level 4 using the sults.
SAA-1 data collected when the excavation was at Levels 2, 3 and 4.
Rmax was the most sensitive small-strain parameter for most cases. 4. Optimization results
Among the large-strain parameters, κ* was the most sensitive parameter
for all cases except for the medium clay layer when optimizing based on Optimizations were carried out for 3 cases when plane strain con-
measurements made at Levels 2, 3 and 4. νpp was the second sensitive ditions applied to the excavation at Excavation Levels 2, 3 and 4.
large-strain parameter, indicating slightly larger sensitivities than the Optimized parameters are compared with HC model parameters re-
λ* parameter. One could expect that κ* was the most sensitive large- ported in literature to assess if the optimized parameters are within a
strain parameter in most of the results because excavations are pri- reasonable range. Optimized parameters are then used to compute the
marily unloading phenomena and κ* represents the unloading/ lateral wall deformations during excavation and are compared to the

5
S. Kim and R.J. Finno Computers and Geotechnics 122 (2020) 103499

Fig. 4. Parameter Correlation Coefficients for all optimization cases of the HC model.

measurements. The initial and optimized HC model parameters are also plotted in
Fig. 5. The range of reported HC model parameters presented in pre-
4.1. Optimized parameters vious studies based on 11 clays [33,39] of 0.09 to 0.38 for νpp and
2.0 × 10−05 to 1.0 × 10−04 for Rmax are indicated in Fig. 5 with the
To help make quantitative assessments of when model parameters dotted lines.
are fully optimized, invariant shear strains (γ) of the soil mass behind Initial and optimized HC model parameters of the medium clay
the sheet pile wall were computed when the excavation reached Levels when optimizing both large-strain and small-strain parameters are il-
2, 3 and 4. The invariant shear strains were computed from results of lustrated as the squares in Fig. 5(a) and (b). Optimized parameters are
the best-fit PLAXIS simulations for each of the excavation levels [17]. within the expected range of parameter values except for the large-
The predominant maximum γ behind the wall in the medium clay in- strain parameter, νpp, optimized at Level 2. Shear strains in the medium
creased with increasing wall deformation and were equal to 0.07, 0.26 clay behind the walls reach only 0.07% at Level 2; this excavation stage
and 0.30% for Levels 2, 3 and 4, respectively. Predominant maximum γ apparently did not provide large enough deformations for reasonable
in the stiff clay were slightly lower: 0.05, 0.12 and 0.20% for Levels 2, 3 optimization of the large-strain parameter. This also may be expected
and 4, respectively. based on the very small sensitivity of the νpp parameter at Level 2
Optimized HC model parameters for all cases are shown in Table 6. shown in Fig. 3, indicating that changes in the large-strain parameter

Table 6
Optimized parameters with optimization stage.
Optimization Stage Both Small, Large Strain Optimized Only Small Strain Optimized

Medium Clay Stiff Clay Medium Clay Stiff Clay

νpp Rmax νpp Rmax Rmax Rmax

Level 2 0.51 5.5 × 10−5 0.12 7.0 × 10−5 5.4 × 10−5 1.9 × 10−4
Levels 2, 3 0.15 3.9 × 10−5 0.09 4.5 × 10−5 3.9 × 10−5 4.6 × 10−5
Levels 2, 3, 4 0.18 3.9 × 10−5 0.19 3.9 × 10−5 3.5 × 10−5 4.2 × 10−5

6
S. Kim and R.J. Finno Computers and Geotechnics 122 (2020) 103499

Fig. 5. Initial and optimized parameters of HC model: (a), (b) Optimizing both small and large strain parameters; (c) Optimizing only small strain parameter.

have very little effects on computations of the lateral deformations at optimize the parameters. In the SQBRC excavation, large incremental
Level 2. When deformations observed at Level 3 are included and the movements that occurred between Levels 2 and 3 precluded a more
shear strain level reaches 0.26% in the medium clay, the optimized νpp precise definition of the strain levels needed for optimization.
decreases from the optimized value based on Level 2, and falls within Initial and optimized HC model parameters for the stiff clay when
the reported range of parameters for various clays. Thereafter, νpp only both small-strain and large-strain parameters are optimized are shown
changes slightly when Level 4 data (maximum shear strain of 0.30%) in Fig. 5(a) and (b) as the triangles. All parameters were within the
are included in the analysis. A similar trend is observed in the opti- reported ranges when both small-strain and large-strain parameters
mized Rmax values of the medium clay. The optimized Rmax value in- were optimized. Changes in the large-strain parameter suggest that
creases during the optimization at Level 2 due to the small deformation optimization was not fully made until the end of excavation at Level 4,
levels. When Level 3 data is included in the optimization, the optimized and larger deformations need to be utilized for the full optimization of
Rmax value decreases and shows little change when Level 4 data is in- the large-strain parameter. The optimized Rmax value at Level 2 in-
cluded. Given the minimal changes in the optimized small-strain and dicated a slight increase due to small deformations measured at that
large-strain parameters between Level 3 and Level 4, it can be con- excavation level. The small deformations used as the basis of the opti-
cluded that the HC model parameters for the medium clay are rea- mization at Level 2 makes the soil represent stiffer responses, thus in-
sonably optimized after Level 3. creasing the optimized Rmax parameter at Level 2. At Level 3, when
Initial and optimized HC model parameters for the medium clay large enough deformations are used as the basis of the optimization, the
when only the small-strain parameter is optimized are illustrated as the optimized Rmax value decreases, indicating a softer response. From the
squares in Fig. 5(c). Note that the large-strain parameter, νpp, is constant little change in the optimized small-strain parameter between optimi-
throughout the optimizations in this case. Optimized small-strain zations of Level 3 and Level 4, it can be concluded that the small-strain
parameters for the medium clay were within the reported range of parameter is fully optimized at Level 3 where shear strains reached
values. When only Level 2 data were considered, Rmax increased slightly 0.12% in the stiff clay stratum. Note that this value falls between the
from the initial values to represent the small deformations used in the 0.07% and 0.26% required for the medium clay.
optimization. The small-strain parameters decreased when Level 3 data Initial and optimized parameters for the stiff clay when only the
were included in the optimization and differed only slightly when Level small-strain parameter is optimized are shown in Fig. 5(c) as the tri-
4 data were utilized, indicating reasonable optimization of the small- angles. When only optimizing the small-strain parameter, Rmax of the
strain parameter of the medium clay after Level 3. stiff clay indicated a large value outside the expected range of para-
For both optimization cases of the medium clay, optimized para- meters at Level 2 due to small deformations measured at that excava-
meters show little changes after Level 3 data are included in the ana- tion level. The optimized parameter decreases when Level 3 data is
lysis. Thus, full optimization of the HC model parameters was complete included and shows little change when Level 4 data is additionally
when the deformations that occur at Level 3 developed. Essentially, the considered, suggesting that the small-strain parameter of the stiff clay
movements were too small when the excavation reached Level 2 to again is fully optimized at Level 3 where shear strains reached 0.12%.
allow the model to be exercised at that stage. Note that the weighting While the stiff clay parameters are not fully optimized when both
function in S(b) includes a measure of the error inherent in the SAA small-strain and large-strain parameters are considered, the small-strain
data. The observations need to be larger than the errors for parameters parameter is fully optimized at Level 3 when only the small-strain
to be reasonably optimized. The HC model parameters were optimized parameter is considered. This suggests that optimizing only the small-
when the excavation reached Level 3 when shear strains of 0.26% oc- strain parameter is more efficient when representing stiff clay behaviors
curred in the soil mass adjacent to the excavation. However, it also can when lateral movements and corresponding soil shear strains are small.
be stated that one needs more than 0.07% to develop in the soil to

7
S. Kim and R.J. Finno Computers and Geotechnics 122 (2020) 103499

Fig. 6. Computations using initial HC model parameters.

4.2. Quality of fit between computed and observed results deformations at the three excavation levels. Note that the boxes in-
dicate the data sets which were used to optimize the parameters. Plots
Fig. 6 shows the measured responses during the excavation and the outside the boxes indicate the “predictions” made by the optimized
computed responses using the initial HC model parameters. It can be parameters.
observed that the computations underestimate the wall responses in the From Fig. 7, a very good fit between measured and computed re-
medium clay and overestimate the deformations in the stiff clay during sponses in the medium and stiff clays at Level 2 can be observed for all
latter excavation stages. However, the computed maximum lateral wall optimization cases, which is expected since the data at Level 2 were
deformation in the medium clay was within 20% of the maximum used to find the optimized parameters. However, the computed “pre-
measured responses at Level 4, not all that bad an estimate of the ob- dictions” of Levels 3 and 4 are significantly smaller than the measured
served responses based on a “prediction.” responses, a consequence of the unreasonable optimization of para-
The degree of fit when optimizing both small-strain and large-strain meters caused by the small deformations at Level 2.
parameters and optimizing only the small-strain parameters is shown in The computed deformations based on inverse analyses of data col-
Figs. 7–9. The figures show the computed and measured lateral wall lected at Excavation Levels 2 and 3 are compared to the observed

Fig. 7. Computations using parameters optimized based on Level 2.

8
S. Kim and R.J. Finno Computers and Geotechnics 122 (2020) 103499

Fig. 8. Computations using parameters optimized based on Levels 2 and 3.

measurements in Fig. 8. Therefore, the results shown for Level 4 in this 1


n

figure represent a “prediction” based on the data collected when the RMSE = (yk yk )2
n (14)
excavation was at Level 3. This “prediction” of Level 4 results in a better
k=1

fit with the observed data than the “prediction” based on optimization
where yk refers to the measured lateral wall deformation at Level k, yk
of only Level 2 data.
indicates the computed lateral wall deformation, and n is the number of
Computed deformations based on inverse analyses of Levels 2, 3 and
observations. Lower RMSE values indicate better fits between the
4 are shown in Fig. 9. In this figure, there is no “prediction,” just a
computations and measurements. RMSE values from the measurements
measure of the fit provided by the inverse analyses using the two
and computations using initial and optimized parameters are shown in
models. Because all three levels of excavation were used in the inverse
Table 7. Note that the RMSE values of Table 7 indicate the degree of fit
analysis, good fits between the computations and measurements are
between the computations and measurements for the excavation levels
observed at all levels.
that were considered in the optimization, not including the “predicted”
To quantify the quality of fit between the measured and computed
computations of latter excavation levels.
lateral wall movements, the Root Mean Squared Error (RMSE) was
As shown in Table 7, RMSE values based on optimizations are
calculated using the following equation:
smaller than the RMSE values computed from the simulation with the

Fig. 9. Computations using parameters optimized based on Levels 2, 3 and 4.

9
S. Kim and R.J. Finno Computers and Geotechnics 122 (2020) 103499

Table 7 Table 8
RMSE values for computations using initial parameters and optimized para- RMSE values for Level 4 “prediction”.
meters from various cases.
Level 4 “Prediction” based on Parameters Small, Large Small Strain
Optimization Stage HC Model from Strain Only

Initial Small, Large Strain Small Strain Only Initial 4.78 4.78
Levels 2, 3 Optimization 4.15 4.13
Level 2 0.45 0.19 0.18
Levels 2, 3 2.85 0.72 0.79
Levels 2, 3, 4 3.61 1.79 2.14 RMSE values that the predictive capabilities slightly improve through
optimization procedures. Differences in predictive capabilities between
optimizing both large-strain and small-strain parameters and opti-
initial HC model parameters, indicating better fits between the com-
mizing only the small-strain parameters are negligible for this excava-
puted and measured deformations after optimizations and showing the
tion.
utility of the inverse analysis method. Optimizing only the small-strain
The maximum lateral wall deformation measured at SAA-1 was
parameter results in slightly larger RMSE values than optimizing both
34 mm at Level 4. Computed maximum lateral wall deformations of
small-strain and large-strain parameters for all optimization cases other
Level 4 using optimized parameters when both small-strain and large-
than the HC model optimization at Level 2, where the 0.01 difference in
strain parameters were optimized and when only the small-strain
RMSE value is negligible. Differences in the RMSE values between op-
parameter was optimized were both 27 mm, with minimal differences
timizing only the small-strain parameter and optimizing both small-
as represented by the very close RMSE values. “Predictions” of the
strain and large-strain parameters are minimal, indicating that opti-
maximum lateral wall deformation in the medium clay were within
mizing only the small-strain parameter does not significantly change
19% of the measured deformations at Level 4. Finno et al. [16] showed
the computed deformations.
that for excavations made in Chicago with good workmanship, that one
can expect to predict excavation-induced lateral wall movements, at
best, to within ± 15% of the maximum value as a consequence of the
4.3. “Predictive” capabilities
natural variability of the Chicago clay. Thus, the simulations presented
herein are as accurate as one could expect.
To quantitatively consider the “predictive” capabilities of the HC
Although the predictions of the maximum lateral wall deformations
model, “Class C predictions” of Level 4 are compared using RMSE va-
were only minimally improved from the initial computations, the
lues. Predictions of Level 4 based on optimizations at Level 2 were not
overall fit of the computed and measured responses in the medium and
considered because the optimized parameters were unreasonable at
stiff clays were significantly improved through the inverse analysis.
Level 2. Only Level 4 computations based on optimizations using Level
2 and 3 observations were considered where the parameters were fully
5. Conclusions
optimized.
Fig. 10 shows the lateral wall deformations computed using the
Soil parameters of the medium and stiff clays in Chicago were op-
initial and optimized parameters and are compared with the measure-
timized based on observed lateral wall deformations at the SQBRC ex-
ments. It can be seen that optimizing both small-strain and large-strain
cavation with inverse analysis. The following conclusions can be drawn
parameters result in negligible differences with results based on opti-
regarding the soil parameter optimization of the HC model:
mizing only the small-strain parameter. However, the degree of fit with
measurements increases substantially compared to the computations
(1) Observed lateral wall deformations were used to conduct sensitivity
made by initial parameters, especially in the stiff stratum.
and correlation analyses, allowing selection of the most effective
Table 8 shows the RMSE values for Level 4 “predictions” based on
parameters to optimize in the SQBRC excavation. When considering
the initial and optimized parameters. It can be seen from the decrease in
both a small-strain parameter and large-strain parameter to opti-
mize, the parameters Rmax and νpp, respectively, were shown to be
the most effective based on the results of the statistical analyses.
Sensitivities of the small-strain parameters were significantly larger
than those of the large-strain parameters due to the relatively small
strains associated with the SQBRC excavation.
(2) When both small-strain and large-strain parameters were opti-
mized, optimized parameters for the medium clay indicated large-
strain parameters outside the reported range of parameters when
using only data from Level 2 due to the small levels of deformation
corresponding to 0.07% shear strain. When additional data from
Level 3 and Level 4 were included in the analysis, optimized
parameters showed insignificant changes in value between Levels 3
and 4. When only the small-strain parameters were optimized, the
optimized parameters of the medium clay were within the reported
range for all excavation stages considered herein. A slight increase
in the small-strain parameter was observed at the optimization for
Level 2 due to the small measured deformation levels and showed
little changes after the Level 3 data were included. HC model
parameters for the medium clay were fully optimized at Level 3.
The shear strain increased from 0.07 to 0.26% as the excavation
was lowered from Level 2 to Level 3, precluding a more accurate
definition of strain level when optimization was complete.
(3) Optimizing both large-strain and small-strain parameters for the
Fig. 10. Computation and measurement at Level 4.

10
S. Kim and R.J. Finno Computers and Geotechnics 122 (2020) 103499

stiff clay did not result in the full optimization of the large-strain [8] Hashash YMA, Levasseur S, Osouli A, Finno RJ, Malecot Y. Comparison of two in-
parameter until the end of excavation due to the small deformations verse analysis techniques for learning deep excavation response. Comput Geotech
2010;37:323–33.
observed in the stiff stratum. When only the small-strain parameter [9] Nikolinakou MA, Whittle AJ, Savidis S, Schran U. Prediction and Interpretation of
was optimized, the small-strain parameter was fully optimized after the performance of a deep excavation in berlin sand. J Geotech Geoenviron Eng
Level 3 data with shear strains of 0.12% were included in inverse 2011;137(11):1047–61.
[10] Mu L, Finno RJ, Huang M, Kim T, Kern K. Defining the soil parameters for com-
analysis, suggesting the efficacy of optimizing only the small-strain puting deformations caused by braced excavation. Maejo Int J Sci Technol
parameter for stiff clays where deformation levels are small. To 2015;9(2):165–80.
reasonably optimize small-strain parameters of stiff clays in ex- [11] Hsiung B, Yang K, Aila W, Hung C. Three-dimensional effects of a deep excavation
on wall deflections in loose to medium dense sands. Comput Geotech
cavations where deformations are small, soil models that consider 2016;80:138–51.
small strain must be used. [12] Peck RB. Deep excavations and tunneling in soft ground. Proc 7th int conf soil mech
(4) RMSE values for optimizations indicated improvements in fit be- found eng; 1969.
[13] Finno RJ. Use of monitoring data to update performance predictions of supported
tween computations and measurements through inverse analysis of
excavations. Proc 7th int symposium on field measurements in geomechanics; 2007.
the HC model. Differences between RMSE values when optimizing [14] Finno RJ. Linking field observations and performance prediction updates during
both small-strain and large-strain parameters and optimizing only construction. Proc 15th great lakes geotech/geoenviron conf; 2008.
the small-strain parameters were minimal. Although “predictions” [15] Whittle AJ, Corral G, Jen LC, Rawnsley RP. Prediction of performance of deep
excavations for Courthouse Station, Boston. J of Geotech Geoenviron Eng
of maximum lateral wall deformations at Level 4 are only mini- 2015;141(4):04014123.
mally improved through optimizations, reaching within 19% of the [16] Finno RJ, Shi Z, Kim S, Crafton N, Rendell D. Adaptive management evaluation of
measured maximum lateral wall deformation, the overall fit be- the SQBRC excavation. Holistic Simulation of Geotechnical Installation Processes.
In: Triantafyllidis T, editor. Lecture notes in applied and computational mechanics.
tween the computed and measured responses improved sub- Springer; 2017. p. 1–28.
stantially. [17] Kim S. Observed performance and inverse analysis of a sheet pile-supported ex-
cavation in Chicago clays PhD thesis Evanston, IL: Northwestern University; 2018.
[18] Kim S, Finno RJ. Inverse analysis of a supported excavation in Chicago. J Geotech
CRediT authorship contribution statement Geoenviron Eng 2019;145(9):04019050.
[19] Finno RJ, Kim S, Lewis J, Van Winkle N. Observed performance of a sheetpile-
Sangrae Kim: Methodology, Software, Formal analysis, supported excavation in Chicago clays. J Geotech Geoenviron Eng
2019;145(2):05018005.
Investigation, Data curation, Writing - original draft, Writing - review & [20] Masin D. Clay hypoplasticity model including stiffness anisotropy. Geotechnique
editing, Visualization. Richard J. Finno: Conceptualization, 2014;64(3):232–8.
Methodology, Supervision, Writing - review & editing, Project admin- [21] Rechea C. Inverse analysis of excavations in urban environments PhD thesis
Evanston, IL: Northwestern University; 2006.
istration, Funding acquisition.
[22] Hill MC. Methods and guidelines for effective model calibration. U.S. Geological
Survey Water-Resources investigations report; 1998, p. 98–4005.
Declaration of Competing Interest [23] Masin D. A hypoplastic constitutive model for clays. Int J Numer Anal Methods
Geomech 2005;29:311–36.
[24] Niemunis A, Herle I. Hypoplastic model for cohesionless soils with elastic strain
The authors declare that they have no known competing financial range. Mech Cohesive-Frictional Mater 1997;2:279–99.
interests or personal relationships that could have appeared to influ- [25] Plaxis B. Reference manual for PLAXIS 2D. Delft, Netherlands: PLAXIS; 2016.
ence the work reported in this paper. [26] Khoiri M, Ou C. Evaluation of deformation parameter for deep excavation in sand
through case histories. Comput Geotech 2013;47:57–67.
[27] Finno RJ, Bryson S, Calvello M. Performance of a stiff support system in soft clay. J
Acknowledgments Geotech Geoenviron Eng 2002;128(8):660–71.
[28] Schanz T, Vermeer PA. On the stiffness of sands. Pre-failure deformation behavior of
geomaterials 1998; 48: 383–387.
The funding for the work reported herein was provided by a [29] Sarabia F. Hypoplastic constitutive law adapted to simulate excavations in Chicago
National Science Foundation grant CMMI-1603060. The support of Dr. glacial clays PhD thesis Evanston, IL: Northwestern University; 2012.
Richard Fragaszy is greatly appreciated. Hayward Baker, Inc. was the [30] Finno RJ, Chung CK. Stress-strain-strength responses of compressible Chicago gla-
cial clays. J Geotech Eng 1992;118(10):1607–25.
excavation support subcontractor. The authors thank Mr. Justin Lewis
[31] Kim T. Incrementally nonlinear responses of soft Chicago glacial clays PhD thesis
of Hayward Baker and Mr. Nathan Van Winkle of GeoEngineers for Evanston, IL: Northwestern University; 2011.
their efforts in collecting the field performance data. [32] Cho W. Recent stress history effects on compressible Chicago glacial clay PhD thesis
Evanston, IL: Northwestern University; 2007.
[33] Arboleda-Monsalve LG. Performance instrumentation and numerical simulation of
References One Museum Park West excavation PhD thesis Evanston, IL: Northwestern
University; 2014.
[1] Finno RJ, Harahap IS. Finite element analyses of HDR-4 excavation. J Geotech Eng [34] Arboleda-Monsalve LG, Teng F, Kim T, Richard RJ. Numerical simulation of triaxial
1991;117(10):1590–609. stress probes and recent stress-history effects of compressible Chicago glacial clays.
[2] Whittle AJ, Hashash YMA, Whitman RV. Analysis of deep excavation in Boston. J J Geotech Geoenviron Eng 2017;143(7):04017129.
Geotech Eng 1993;119(1):69–90. [35] Kim T, Finno RJ. Anisotropy evaluation and irrecoverable deformation in triaxial
[3] Ou CY, Tang YG. Soil parameter determination for deep excavation analysis by stress probes. J Geotech Geoenviron Eng 2012;138(2):155–65.
optimization. J Chin Inst Eng 1994;17(5):671–88. [36] Finno RJ, Hiltunen DR, Kim T. Elastic cross-anisotropy of Chicago glacial clays from
[4] Hashash YMA, Whittle A. Ground movement prediction for deep excavations in soft field and laboratory data. In: Proc GeoCongress 2012: state of the art and practice in
clay. J Geotech Eng 1996;122(6):474–86. geotech eng.
[5] Finno RJ, Calvello M. Supported excavations: observational method and inverse [37] Masin D, Rott J. Small strain stiffness anisotropy of natural sedimentary clays: re-
modeling. J Geotech Geoenviron Eng 2005;131(7):826–36. view and a model. Acta Geotech 2014;9(2):299–312.
[6] Finno RJ, Roboski JF. Three-dimensional responses of a tied-back excavation [38] Masin D. Hypoplasticity for practical applications PhD thesis Prague, Czech
through clay. J Geotech Geoenviron Eng 2005;131(3):273–82. Republic: Charles University; 2011.
[7] Blackburn JT, Finno RJ. Three-dimensional responses observed in an internally [39] Gudehus G, Amorosi A, Gens A, Herle I, Kolymbas D, Masin D, et al. The soilmo-
braced excavation in soft clay. J Geotech Geoenviron Eng 2007;133(11):1364–73. dels.info project. Int J Numer Anal Methods Geomech 2008;32(12):1571–2.

11

You might also like